XGBoost GPU Support

This page contains information about GPU algorithms supported in XGBoost.

Note

CUDA 10.1, Compute Capability 3.5 required

The GPU algorithms in XGBoost require a graphics card with compute capability 3.5 or higher, with CUDA toolkits 10.1 or later. (See this list to look up compute capability of your GPU card.)

CUDA Accelerated Tree Construction Algorithms

Tree construction (training) and prediction can be accelerated with CUDA-capable GPUs.

Usage

Specify the tree_method parameter as one of the following algorithms.

Algorithms

tree_method

Description

gpu_hist

Equivalent to the XGBoost fast histogram algorithm. Much faster and uses considerably less memory. NOTE: May run very slowly on GPUs older than Pascal architecture.

Supported parameters

GPU accelerated prediction is enabled by default for the above mentioned tree_method parameters but can be switched to CPU prediction by setting predictor to cpu_predictor. This could be useful if you want to conserve GPU memory. Likewise when using CPU algorithms, GPU accelerated prediction can be enabled by setting predictor to gpu_predictor.

The device ordinal (which GPU to use if you have many of them) can be selected using the gpu_id parameter, which defaults to 0 (the first device reported by CUDA runtime).

The GPU algorithms currently work with CLI, Python, R, and JVM packages. See Installation Guide for details.

Python example
param['gpu_id'] = 0
param['tree_method'] = 'gpu_hist'
With Scikit-Learn interface
XGBRegressor(tree_method='gpu_hist', gpu_id=0)

GPU-Accelerated SHAP values

XGBoost makes use of GPUTreeShap as a backend for computing shap values when the GPU predictor is selected.

model.set_param({"predictor": "gpu_predictor"})
shap_values = model.predict(dtrain, pred_contribs=True)
shap_interaction_values = model.predict(dtrain, pred_interactions=True)

See examples here.

Multi-node Multi-GPU Training

XGBoost supports fully distributed GPU training using Dask. For getting started see our tutorial Distributed XGBoost with Dask and worked examples here, also Python documentation Dask API for complete reference.

Objective functions

Most of the objective functions implemented in XGBoost can be run on GPU. Following table shows current support status.

Objectives

GPU support

reg:squarederror

reg:squaredlogerror

reg:logistic

reg:pseudohubererror

binary:logistic

binary:logitraw

binary:hinge

count:poisson

reg:gamma

reg:tweedie

multi:softmax

multi:softprob

survival:cox

survival:aft

rank:pairwise

rank:ndcg

rank:map

Objective will run on GPU if GPU updater (gpu_hist), otherwise they will run on CPU by default. For unsupported objectives XGBoost will fall back to using CPU implementation by default. Note that when using GPU ranking objective, the result is not deterministic due to the non-associative aspect of floating point summation.

Metric functions

Following table shows current support status for evaluation metrics on the GPU.

Metric

GPU Support

rmse

rmsle

mae

mape

mphe

logloss

error

merror

mlogloss

auc

aucpr

ndcg

map

poisson-nloglik

gamma-nloglik

cox-nloglik

aft-nloglik

interval-regression-accuracy

gamma-deviance

tweedie-nloglik

Similar to objective functions, default device for metrics is selected based on tree updater and predictor (which is selected based on tree updater).

Benchmarks

You can run benchmarks on synthetic data for binary classification:

python tests/benchmark/benchmark_tree.py --tree_method=gpu_hist
python tests/benchmark/benchmark_tree.py --tree_method=hist

Training time on 1,000,000 rows x 50 columns of random data with 500 boosting iterations and 0.25/0.75 test/train split with AMD Ryzen 7 2700 8 core @3.20GHz and NVIDIA 1080ti yields the following results:

tree_method

Time (s)

gpu_hist

12.57

hist

36.01

Memory usage

The following are some guidelines on the device memory usage of the gpu_hist tree method.

Memory inside xgboost training is generally allocated for two reasons - storing the dataset and working memory.

The dataset itself is stored on device in a compressed ELLPACK format. The ELLPACK format is a type of sparse matrix that stores elements with a constant row stride. This format is convenient for parallel computation when compared to CSR because the row index of each element is known directly from its address in memory. The disadvantage of the ELLPACK format is that it becomes less memory efficient if the maximum row length is significantly more than the average row length. Elements are quantised and stored as integers. These integers are compressed to a minimum bit length. Depending on the number of features, we usually don’t need the full range of a 32 bit integer to store elements and so compress this down. The compressed, quantised ELLPACK format will commonly use 1/4 the space of a CSR matrix stored in floating point.

Working memory is allocated inside the algorithm proportional to the number of rows to keep track of gradients, tree positions and other per row statistics. Memory is allocated for histogram bins proportional to the number of bins, number of features and nodes in the tree. For performance reasons we keep histograms in memory from previous nodes in the tree, when a certain threshold of memory usage is passed we stop doing this to conserve memory at some performance loss.

If you are getting out-of-memory errors on a big dataset, try the or xgboost.DeviceQuantileDMatrix or external memory version.

Developer notes

The application may be profiled with annotations by specifying USE_NTVX to cmake. Regions covered by the ‘Monitor’ class in CUDA code will automatically appear in the nsight profiler when verbosity is set to 3.

References

Mitchell R, Frank E. (2017) Accelerating the XGBoost algorithm using GPU computing. PeerJ Computer Science 3:e127 https://doi.org/10.7717/peerj-cs.127

NVIDIA Parallel Forall: Gradient Boosting, Decision Trees and XGBoost with CUDA

Out-of-Core GPU Gradient Boosting

Contributors

Many thanks to the following contributors (alphabetical order):

  • Andrey Adinets

  • Jiaming Yuan

  • Jonathan C. McKinney

  • Matthew Jones

  • Philip Cho

  • Rong Ou

  • Rory Mitchell

  • Shankara Rao Thejaswi Nanditale

  • Sriram Chandramouli

  • Vinay Deshpande

Please report bugs to the XGBoost issues list: https://github.com/dmlc/xgboost/issues. For general questions please visit our user form: https://discuss.xgboost.ai/.